Introduction

This section will be written by the pipeline.

Cardin et al., 2009Sohal et al., 2009Optogenetic studies in mice indicate that activating fast-spiking or parvalbumin interneurons can amplify or generate cortical gamma rhythms, while gamma-cycle timing can shape sensory-response precision.

Machine learning discovered, painfully, that undocumented datasets are a primary source of downstream harm. Gebru et al., 2021Gebru and colleagues proposed datasheets for datasets: a standard document recording a dataset’s motivation, composition, collection process, recommended uses, and maintenance - modelled on the datasheets that accompany electronic components. Their key insight for us is not the template but its function: documentation is a debt-prevention instrument, one that makes deferred decisions visible at the point of reuse, forcing the creator to confront their collection assumptions and letting the consumer judge fitness for a new purpose.

Mitchell et al., 2019Mitchell and colleagues added model cards: short reports of a model’s intended use, its limitations, and - crucially - its performance broken out by subgroup and context rather than as a single headline number. Disaggregated reporting surfaces exactly what an aggregate conceals.

Bender & Friedman, 2018Bender and Friedman proposed data statements for language technology to document represented populations and dataset provenance; they argued that this practice could reduce bias and improve the precision of claims about generalization.